fix(dataset): raise a clear error when custom dataset_info is not a list

register_dataset_info loads the --custom_dataset_info JSON and then does
`for d_info in dataset_info: _register_d_info(d_info)`. When a user writes a
single entry as a top-level object `{...}` instead of the documented one-element
list `[{...}]`, `for d_info in dataset_info` iterates the dict keys (strings),
and _preprocess_d_info then does `d_info['preprocess_func'] = ...` on a str,
raising a cryptic `TypeError: 'str' object does not support item assignment`.
The built-in dataset_info.json is a list and the docs state it "is a list of
metadata for datasets", so a non-list is a user mistake that deserves a clear
message.

Validate the loaded dataset_info is a list and raise a helpful ValueError
otherwise. Valid list inputs are unaffected.

Author: he-yufeng

swift/dataset/register.py

@@ -103,6 +103,10 @@ def register_dataset_info(dataset_info: Union[str, List[str], None] = None) -> L
                 dataset_info = json.load(f)
         else:
             dataset_info = json.loads(dataset_info)  # json
+    if not isinstance(dataset_info, list):
+        raise ValueError(
+            f'`dataset_info` should be a JSON list of dataset entries (e.g. `[{{...}}]`), but got '
+            f'type `{type(dataset_info).__name__}`. Please wrap the entry in a list.')
     if len(dataset_info) == 0:
         return []
     res = []

tests/general/test_dataset.py

@@ -79,6 +79,27 @@ def test_cls():
     _test_dataset(['simpleai/HC3-Chinese:baike_cls'])
 
 
+def test_register_dataset_info_rejects_non_list():
+    import json
+    import tempfile
+
+    from swift.dataset import register_dataset_info
+
+    # A single entry written as a top-level object instead of a one-element list is a
+    # common mistake; it must fail with a clear message rather than a cryptic
+    # "'str' object does not support item assignment" from iterating the dict keys.
+    with tempfile.NamedTemporaryFile('w', suffix='.json', delete=False, encoding='utf-8') as f:
+        json.dump({'ms_dataset_id': 'swift/self-cognition'}, f)
+        path = f.name
+
+    try:
+        register_dataset_info(path)
+    except ValueError as e:
+        assert 'list' in str(e)
+    else:
+        raise AssertionError('expected ValueError for a non-list dataset_info')
+
+
 if __name__ == '__main__':
     # test_sft()
     # test_agent()